Construction of speech acquisition mechanism based on sensory information

我们在语音清晰度感知反馈的扰动研究和利用机械式说话机器人模仿人类语音产生的建构性研究中，研究了基于感知信息的语音习得机制.扰动研究：我们使用扰动范式研究感觉反馈对言语产生的影响。在掩蔽听觉和触觉反馈条件下，外部扰动意外地作用于语音发声过程中的腭部形状。然后，我们通过使用功能磁共振成像观察大脑活动的补偿过程中的舌头的扰动。我们发现这种代偿反应是由中枢神经系统的神经适应机制引起的.会说话的机器人研究：我们已经开发了一个三维说话机器人（机械语音合成器），模仿人类的语音产生过程。该机器人具有人工机械肺、声门、舌头、下颚、嘴唇和软腭作为发音器官 关于我们 其由伺服机构动态控制。语音是由自振荡声带产生的声门源、气流通过声道狭窄处产生的随机源和时变声道声共振产生的。采用前馈-反馈机制对说话机器人进行动态控制。控制规划是针对期望的语音声学和感觉信息（如舌头与腭之间的触觉信息和口腔压力）自动训练的。提出了一种基于隐马尔可夫模型的随机语音反演方法。此外，我们提出了一个规划方法的基础上的一个更新规则，以确定发音位置的感官信息。喉控制规划也研究了基于前-逆神经网络模型确定机器人参数喉为所需的声学，如音调频率，清音信息，和语音质量。少

We have been studied speech acquisition mechanism based on sensory information in perturbation studies of sensory feedback during speech articulation and constructive studies by using mechanical talking robot mimicking human speech production.1. Perturbation Study : We investigated sensory feedback effects on speech production by using a perturbation paradigm. The external perturbation was unexpectedly applied on the palatal shape during speech utterance under the masked auditory and tactile feedback conditions. Then, we observed brain activity by using fMRI during the compensatory response by the tongue to the perturbation. We found that the compensatory response was brought by the neural adaptation mechanism in the central nerves system.2. Talking Robot Study : We have developed an 3 dimensional talking robot (mechanical speech synthesizer) which mimicking human speech production process. This robot has artificial mechanical lung, glottis, tongue, jaw, lips and velum as speech organs … More which are dynamically controlled by servo mechanism. The speech sounds are produced by glottal source generated by self-oscillating vocal cords, random source generated by passing the air flow through the narrow constriction in the vocal tract, and time varying vocal tract acoustic resonance. The talking robot is dynamically controlled by feed-forward-feedback mechanism. The control planning is automatically trained for a desired speech acoustics and sensory information like tactile information between the tongue and the palate and oral pressure. We proposed a stochastic speech inversion method for mapping the speech acoustics to the articulatory positions based on Hidden Markov Model. Also, we proposed a planning method based on a Jacobean updating rule to determine the articulatory positions from the sensory information. The larynx control planning is also studied based on forward-inverse neural network model for determining the robot parameters of larynx for a desired acoustics like pitch frequency, voice-unvoiced information, and voice quality. Less